Embed Size (px)
Citation preview
Manil Maskey, Dan Cecil, Rahul Ramachandran, and JJ MillerNASA Marshall Space Flight Center
Tropical Cyclone Intensity Estimation Using Deep Convolutional Neural Networks
33rd Conference on Hurricanes and Tropical MeteorologyApril 20, 2018. Pointe Vedra, FL
https://ntrs.nasa.gov/search.jsp?R=20180002953 2020-03-13T03:48:02+00:00Z
Overview
• Deep learning and Convolutional Neural Network
• CNN for Tropical Cyclone Intensity Estimation
• Preliminary results
• Work in progress
Deep Learning
• A subfield of machine learning
• Algorithms inspired by function of the brain
• Scales with amount of training data
• Powerful tool without the need for feature engineering
• Suitable for many Earth Science applications
Traditional Image Classification Approach
• Image Features: Color, Texture, Edge histogram,…
• “Shallow” architecture
• Experts define features
Hand-craftedFeature Extractor
“Simple”Trainable Classifier
(static) (learns)
“DEEP” Architecture• Features are key to recognition
• What about learning the features?
• Deep Learning• Hierarchical Learning
• Modeled after human brain
• Process information through multiple stages of transformation and representation
Hand-craftedFeature Extractor
“Simple”Trainable Classifier
(static) (learns)
Cost(Compared using
Labeled Data)
Back propagation
Classification
Convolutional Neural Network• Input image – labeled training data• Convolution Layers – filters are applied across input
images (start with random filters)• Non-linearity – a bias function so that the network
is not remembering but rather generalizing• Pooling – subsampling of the output so that the
images do not grow exponentially• Final output images are passed through a
traditional neural network for classification• Classification results are compared using a loss
function to determine error• Based on error the weights and filters are adjusted
using gradient descent• Iterate the process until the error is below some
threshold
0 0 0 0 0 0 0
0 1 2 1 0 1 0
0 2 1 1 2 1 0
0 1 2 1 2 2 0
0 2 2 2 1 0 0
0 0 1 1 1 2 0
0 0 0 0 0 0 0
1 1 -1
0 -1 1
-1 0 1
2 0 -3
4 0 0
1 3 -1
Input (7x7), pad of 1 Kernel (3x3), stride of 2 Output (3x3)
2D Convolution, Single Slice
• Stride (s)• Jump/step with which filters move across width/height of input volume
• Padding (p)• Amount of wrapping used in input
• Output size (Wo) = (Wi – k + 2*p)/s + 1
Convolutional Layer
Network architecture
• The Dvorak technique• Vernon Dvorak (1970s)• Satellite-based method• Cloud system measurements• Development patterns corresponds to T-number
• Deviation-angle variation technique (DAVT)• Piñeros et al. (2008)• Variance for quantification of cyclones• Calculates using center (eye) pixel• Directional gradient statistical analysis of the brightness of images
Tropical Cyclone Intensity Estimation
Issues
• Subjective/Uncertainty
• Lack of generalizability
• Inconsistency
• Complexity
Can we objectively predict wind speed from images?
• Images• US Naval Research Laboratory (http://www.nrlmry.navy.mil/tcdat)• From 1998 to 2014• Images at 15 minutes interval
• Cyclone data• National Hurricane Center (http://www.nhc.noaa.gov) (HURDAT and HURDAT2)• Hurricane Research Division (http://www.aoml.noaa.gov/hrd/hurdat/Data_Storm.html)• Every 6 hours
• 98 cyclones collected over Pacific and Atlantic regions• 68 from Atlantic• 30 from Pacific
Data
Storms
• Interpolate to increase even more• 2 hours interpolated image differences
2 hour interpolated image differences
Data augmentation
• (Training + Validation) 70% - 30% (Test)• (Training) 75% - 25% (Validation)
Training, test, and validation
Feature maps from second convolution
Visualization
• Model with around 90% of validation accuracy• Tested against 14,345 test images (Atlantic + Pacific)
• Confusion Matrix• Classification Report• Accuracy• RMS Intensity Error
Initial performance
• Top-1: exact-hits• Top-2: exact-hits + 2nd-hits
Accuracy
• Our model• Across Atlantic and Pacific• Achieved RMSE of 9.19kt
• North Atlantic• Piñeros et al. (2011): 14.7kt • Ritchie et al. (2012): 12.9kt
• North Pacific• Ritchie et al. (2014): 14.3kt
Error Metrics
Sample correct classifications
Sample incorrect classifications
Adapted from Stevenson et al. (2014). Time series of satellite-derived intensity estimates (circles) for Hurricane Earl (2010), added to best track intensities and lightning flash rate time series.
Detailed look: Hurricane Earl, 2010
• Hurricane intensity estimation portal• Use of passive microwave dataset• Use of atmospheric conditions
Work in progress
Hurricane intensity estimation portal• Develop a near real-time tropical cyclone intensity estimation services
• Include additional image datasets• Algorithmic enhancements• Monitor NHC outlook for “invest” area for trigger
• Perform extensive evaluation with available observations• Work with NASA/SPoRT to develop a website that will display current “invest” information along with
estimated wind speed information and relevant overlays • Develop OGC services (WFS and SOS): integration with AWIPS/N-AWIPS
Hurricane intensity estimation portal
http://hiep.surge.sh/storms/9eee5297-d43d-4f84-9931-23bef5fbdbb4
Thank you.
Using Microwave Datasets
Instrument (85, 89 GHz) Coverare years Total storm centric images
SSMI17 2008-2016 1715
SSMI18 2010-2016 1378
TMI 1998-2014 3409
AMSRE 2003-2011 2230
Network
CNN for GOES images.
CNN for Microwave Images
GOES images
(616 X 616)
MW Images
(366 X 366)
Merge layer
Dense layer
Dense layer
Categories
Process1. Collect Storm-centric PM data2. Generate image3. Match up images with NRL goes images4. Add random rotation/flips to images (data augmentation).5. Use corresponding GOES and Microwave images for training.6. Start with 7 categories (ts, td, 1, 2, 3, 4, 5)
Samples
Source: AMSRE/GOESWind speed: 145Hurricane: DeanYear: 2007
Source: TMI/GOESWind speed: 125Hurricane: DeanYear: 2007
Source: SSMI18/GOESWind speed: 125Hurricane: MatthewYear: 2016
